U.S. patent application number 12/315657 was filed with the patent office on 2010-06-10 for designing minimum total cost networks using iterative rounding approximation methods.
This patent application is currently assigned to AT&T Intellectual Property I, L.P. Invention is credited to Mohammad Hajiaghayi.
Application Number | 20100142408 12/315657 |
Document ID | / |
Family ID | 42230956 |
Filed Date | 2010-06-10 |
United States Patent
Application |
20100142408 |
Kind Code |
A1 |
Hajiaghayi; Mohammad |
June 10, 2010 |
Designing minimum total cost networks using iterative rounding
approximation methods
Abstract
Minimum cost networks, such as fiber optic networks used in
telecommunications, are obtained by defining available network
elements having cost, required pairs, connectivity and penalty cost
values and selecting from these available elements using an
iterative rounding approximation method that constructs an LP
relaxation incorporating the element parameters, finds an optimal
basic solution, applies a selection criterion to pairs and edges in
the optimal basic solution, and constructs a residual LP relaxation
with selected pairs and edges. By fixing selected pairs and edges
values to 1 in the residual LP, successive iterations of the method
provide a design which is a 3-approximation solution to the minimum
cost design problem.
Inventors: |
Hajiaghayi; Mohammad;
(Florham Park, NJ) |
Correspondence
Address: |
AT & T Legal Department - Lazaroff
Attn: Patent Docketing, Room 2A-207, One AT & T Way
Bedminster
NJ
07921
US
|
Assignee: |
AT&T Intellectual Property I,
L.P
|
Family ID: |
42230956 |
Appl. No.: |
12/315657 |
Filed: |
December 5, 2008 |
Current U.S.
Class: |
370/255 |
Current CPC
Class: |
H04L 41/145 20130101;
H04L 45/12 20130101 |
Class at
Publication: |
370/255 |
International
Class: |
H04L 12/28 20060101
H04L012/28 |
Claims
1. An approximation method for designing a minimum cost network
from defined available network elements having values for cost,
required pairs, connectivity requirements and penalty costs for
insufficient connectivity by selecting from these available
elements, comprising: constructing an LP relaxation for a graph
having edges and vertices with values corresponding to those of the
available network elements; finding an optimal basic solution to
the LP; applying a selection criterion to penalty costs and edge
costs identified in the optimal basic solution and including the
selected penalty costs and edges in the minimum cost network;
constructing a residual LP relaxation including remaining pairs
with unsatisfied connectivity and edges; and fixing selected edge
coefficient values to 1 in the residual LP; repeating the steps of
the method to include additional penalty costs and edge costs in
the minimum cost network until all connectivity requirements are
satisfied.
2. An approximation method as claimed in claim 1 wherein the step
of applying a selection criterion comprises selecting a number a,
and including all penalty costs and edges in the minimum cost
network that are identified in the optimal basic solution with
coefficients.gtoreq..alpha..
3. An approximation method as claimed in claim 1 wherein
.alpha.=1/3.
4. An approximation method as claimed in claim 1 wherein the step
of applying a selection criterion comprises including the penalty
cost and edge in the minimum cost network that are identified in
the optimal basic solution with the maximum coefficient values.
5. An approximation method as claimed in claim 1 wherein the
connectivity requirement for each required pair is for a single
connection between the vertices of the pair.
6. An approximation method as claimed in claim 1 wherein the
connectivity requirement for at least one required pair is for a
plurality of independent connections between the vertices of the
pair.
7. An approximation method as claimed in claim 1 wherein the method
provides a 3-approximation algorithmic solution.
8. An approximation method as claimed in claim 1 further
comprising: selecting from the available network elements at least
some that correspond to edges selected for inclusion in the minimum
cost network.
9. An approximation method as claimed in claim 1 wherein the
minimum cost network to be designed is a telecommunications
network, and wherein the network elements include transmission
links.
10. An approximation method as claimed in claim 9 wherein the
transmission links are fiber optic transmission links.
11. An approximation method as claimed in claim 1 wherein the
minimum cost network to be designed is a package delivery network,
and wherein the network elements include transportation links.
12. An approximation method as claimed in claim 1 wherein the
minimum cost network to be designed is a telecommunications
network, wherein the network elements include transmission links
with construction costs, the required pairs are pairs of nodes
required to be connected in the network, the connectivity
requirements set the number of independent paths between the nodes
of a pair that the network is required to supply, and the penalty
costs are the costs of buying transmission capacity from commercial
sources to form paths between nodes of a required pair to satisfy
connectivity requirements.
13. An approximation method as claimed in claim 1 wherein the
method steps are carried out by computer readable program code
operable by a computer system.
14. An approximation method for designing a minimum total cost
network by selecting from a collection of available network
elements which include nodes and links connecting the nodes,
comprising: selecting graph edges to correspond to the available
links; assigning edge costs that correspond to the link costs;
selecting graph vertices to correspond to nodes; selecting pairs of
vertices that are required to be in the network; assigning
connectivity requirements for all required pairs of vertices;
assigning penalty costs to the required pairs if their connectivity
requirements are not satisfied; constructing an LP relaxation
incorporating values for edge costs, pair requirements,
connectivity requirements and penalty costs and having coefficients
for edge costs and penalty costs as variables; finding an optimal
basic feasible solution to the LP; applying a selection criterion
to pair connectivities and edges in the optimal basic solution;
paying a penalty cost for each pair connectivity selected according
to the selection criterion and including the penalty cost in the
network solution; including each edge selected according to the
selection criterion in the network solution; forming a residual
graph including all remaining required pairs with unpaid
connectivity requirements and including all edges; constructing a
residual LP for the residual graph in which edge coefficients for
edges selected according to the selection criterion are set to 1
and other edge coefficients are variables; iteratively repeating
the previous steps from the finding step to the constructing step
until all connectivity requirements of the required pairs are
determined to be satisfied; and selecting network links
corresponding to at least some edges identified in the network
solution to include in the minimum cost network.
15. An approximation method as claimed in claim 14 wherein the step
of applying a selection criterion comprises selecting a number a,
and including all penalty costs and edges in the minimum cost
network that are identified in the optimal basic solution with
coefficients.gtoreq..alpha..
16. An approximation method as claimed in claim 15 wherein
.alpha.=1/3.
17. An approximation method as claimed in claim 14 wherein the step
of applying a selection criterion comprises including the penalty
cost and edge in the minimum cost network that are identified in
the optimal basic solution with the maximum coefficient values.
18. An approximation method as claimed in claim 14 wherein the
connectivity requirement for each required pair is for a single
connection between the vertices of the pair.
19. An approximation method as claimed in claim 14 wherein the
connectivity requirement for at least one required pair is for a
plurality of independent connections between the vertices of the
pair.
20. A system for designing a minimum cost network from defined
available network elements having values for cost, required pairs,
connectivity requirements and penalty costs for insufficient
connectivity by selecting from these available elements,
comprising: means for providing an LP relaxation for a graph having
edges and vertices with values corresponding to those of the
available network elements; means for finding an optimal basic
solution to the LP; means for applying a selection criterion to
penalty costs and edge costs identified in the optimal basic
solution and including the selected penalty costs and edges in the
minimum cost network; means for providing a residual LP relaxation
including remaining pairs with unsatisfied connectivity and edges;
and means for fixing selected edge coefficient values to 1 in the
residual LP; means for repeating the steps of the method to include
additional penalty costs and edge costs in the minimum cost network
until all connectivity requirements are satisfied.
21. A system as claimed in claim 20 wherein the means for applying
a selection criterion comprises means for selecting a number a, and
means for including all penalty costs and edges in the minimum cost
network that are identified in the optimal basic solution with
coefficients.gtoreq..alpha..
22. A system as claimed in claim 21 wherein .alpha.=1/3.
23. A system as claimed in claim 20 wherein the means for applying
a selection criterion comprises means for including the penalty
cost and edge in the minimum cost network that are identified in
the optimal basic solution with the maximum coefficient values.
24. A computer program product for causing a computer to design a
minimum total cost network from defined available network elements
having values for cost, required pairs, connectivity requirements
and penalty costs for insufficient connectivity by selecting from
these available elements, comprising: computer readable program
code means for providing an LP relaxation for a graph having edges
and vertices with values corresponding to those of the available
network elements; computer readable program code means for finding
an optimal basic solution to the LP; computer readable program code
means for applying a selection criterion to penalty costs and edge
costs identified in the optimal basic solution and including the
selected penalty costs and edges in the minimum cost network;
computer readable program code means for providing a residual LP
relaxation including remaining pairs with unsatisfied connectivity
and edges; and computer readable program code means for fixing
selected edge coefficient values to 1 in the residual LP; computer
readable program code means for repeating the steps of the method
to include additional penalty costs and edge costs in the minimum
cost network until all connectivity requirements are satisfied.
25. A computer program product as claimed in claim 24 wherein the
computer readable program code means for applying a selection
criterion comprises computer readable program code means for
selecting a number a, and computer readable program code means for
including all penalty costs and edges in the minimum cost network
that are identified in the optimal basic solution with
coefficients.gtoreq..alpha..
26. A computer program product as claimed in claim 25 wherein
.alpha.=1/3.
27. A computer program product as claimed in claim 24 wherein the
computer readable program code means for applying a selection
criterion comprises computer readable program code means for
including the penalty cost and edge in the minimum cost network
that are identified in the optimal basic solution with the maximum
coefficient values.
28. A minimum cost network assembled from defined available network
elements having values for cost, required pairs, connectivity
requirements and penalty costs for insufficient connectivity by
selecting from these available elements using a method comprising:
constructing an LP relaxation for a graph having edges and vertices
with values corresponding to those of the available network
elements; finding an optimal basic solution to the LP; applying a
selection criterion to penalty costs and edge costs identified in
the optimal basic solution and including the selected penalty costs
and edges in the minimum cost network; constructing a residual LP
relaxation including remaining pairs with unsatisfied connectivity
and edges; and fixing selected edge coefficient values to 1 in the
residual LP; repeating the steps of the method to include
additional penalty costs and edge costs in the minimum cost network
until all connectivity requirements are satisfied; selecting from
the available network elements at least some that correspond to
edges selected for inclusion in the minimum cost network.
Description
FIELD OF THE INVENTION
[0001] The present invention generally relates to network design,
and more particularly to a method and system for designing minimum
total cost networks, to a computer program product for performing
such methods, and to networks designed according to such
methods.
BACKGROUND OF THE INVENTION
[0002] Network design applies to many fields of commercial
interest. One example is overnight package delivery, where packages
must be delivered between cities over a network of possible air
routes that have different costs. Another example is utility
design, where a commodity such as electricity or water is to be
distributed to geographically separate customers over a network of
possible conduits that have different costs. The field of
telecommunications contains several examples in which possible
transmission elements, with different costs, may be combined in
many different ways, between different points, in order to obtain a
network of paths between the points. One example arises in the
design of a fiber optic transmission network (e.g., one using fiber
optic links between routers as part of an ISP network) to join
geographically dispersed major cities, with possible alternate
routes between cities having different costs (which may be defined
in terms of the construction costs of providing the route, for
example).
[0003] In each of these network examples, it is desirable to be
able to employ a method or algorithm which will provide the minimum
cost network by selecting, from the possible network elements,
those which provide the lowest total cost. A minimum cost network
designed according to such a method will allow lower cost services
to customers using such a network and give a competitive advantage
to the service provider using such a network.
[0004] The overall network design context may include additional
considerations. For example, the network design criteria may
require that the network include pairs of selected points, e.g.,
certain city pairs. The design problem then is to find the minimum
cost network solution that contains paths between each of these
paired points.
[0005] The network design criteria may further assign penalty costs
if a required path between selected points is not included. An
example of such a penalty cost in the package delivery context is
the cost of leasing another carrier's air transport facilities for
a package delivery route segment not covered by the package
deliverer's own network. In the telecommunications context, an
example of such a penalty cost is the cost of leasing
telecommunications transmission capacity from another provider for
a route segment not provided by a provider's own network. Allowing
penalty costs supplies the opportunity for a lower overall cost
network. The network design problem, with this additional
consideration, is then to find a network solution that minimizes
the total combined costs of selected network elements plus
penalties incurred for having any required points unconnected by
the selected elements.
[0006] Finding a network whose design considerations include a
minimization of cost, the selection of elements with assigned
costs, the requirement that the solution include pairs of selected
points, and the allowance of assigned penalty costs if a required
pair of points is not connected by the selected elements, is a
design problem corresponding mathematically to the Prize Collecting
Steiner Forest (PCSF) problem. Mathematically, the network of
possible elements is a graph G=(V,E) where V is the set of vertices
v.sub.i, v.sub.j, . . . , and E is the set of edges e.sub.ij
connecting vertices v.sub.i, v.sub.j, with a set of required pairs
={(s.sub.1,t.sub.1), (s.sub.2,t.sub.2), . . . (st}, a non-negative
cost function c: E.fwdarw.Q+ (which means the function maps the set
E into a set of positive rational numbers), and a non-negative
penalty function .pi.: Q+. Each edge has a cost c.sub.ij and each
pair has a penalty .pi..sub.k that is incurred if the pair vertices
s.sub.k, t.sub.k are not connected to an edge. The goal is to find
a set (or forest) H that minimizes the costs of the edges of H plus
the penalties paid for pairs whose vertices are not all connected
by H. In other words, the goal is a minimum-cost way of buying a
set of edges and paying the penalty for those pairs which are not
connected via bought edges.
[0007] A simple example of such a network is shown in FIG. 1. The
graph G of potential network elements includes hollow circle
vertices that are denoted with upper case letters A, B, C, D, F and
J and represent cities, for example. The solid circles denoted with
lower case letters a, b, c, d and f represent non-city vertices.
Edges e representing transmission links, for example, that are
available to be included in the network, are denoted by solid lines
linking vertices as shown. Each edge e.sub.ij has a cost c.sub.ij.
Required pairs of vertices , , , , representing cities that are
required to be connected in the network, have respective penalty
costs .pi..sub.1, .pi..sub.2, .pi..sub.3, .pi..sub.4, that are
incurred if the vertices in the pair are not joined to a selected
edge.
[0008] A minimal cost solution H to the graph G of FIG. 1 is shown
in FIG. 1A, where solid lines represent edges selected from those
available in FIG. 1, and dashed lines show penalty connections. The
connection for required pair AB is furnished by penalty .pi..sub.1;
for pair BC by edges Bc, cD; for pair CD by edge CD; and for pair
DF by edges DC, Cc, cJ and JF.
[0009] The Prize Collecting Steiner Forest (PCSF) problem has been
solved for special cases. In the special case that all sinks are
identical, i.e., there is a common root r in every pair in the PCSF
problem, the problem reduces to the classic Prize-Collecting
Steiner Tree problem. Bienstock, Goemans, Simchi-Levi, and
Williamson [D. BIENSTOCK, M. X. GOEMANS, D. SIMCHI-LEVI, AND D.
WILLIAMSON, A note on the prize collecting traveling salesman
problem, Math. Programming, 59 (1993), pp. 413-420] first
considered this tree problem (based on a problem earlier proposed
by Balas [E. BALAS, The prize collecting traveling salesman
problem, Networks, 19 (1989), pp. 621-636]) for which they gave a
3-approximation algorithm. The current best approximation algorithm
for this tree problem is a primal-dual 2-1/(n-1) approximation
algorithm (n is the number of vertices of the graph) due to Goemans
and Williamson [M. X. GOEMANS AND D. P. WILLIAMSON, A general
approximation technique for constrained forest problems, SIAM J.
Comput., 24 (1995), pp. 296-317].
[0010] The general form of the PCSF problem first was formulated by
Hajiaghayi and Jain [M. T. HAJIAGHAYI AND K. JAIN, The
prize-collecting generalized Steiner tree problem via a new
approach of primal-dual schema, in Proceedings of the Seventeenth
Annual ACM-SIAM Symposium on Discrete Algorithms, New York, 2006,
ACM, pp. 631-640]. They showed how by a primal-dual algorithm to a
novel integer programming formulation of the problem with
doubly-exponential variables, it is possible to obtain a
3-approximation algorithm for the problem. In addition, they show
that the factor 3 in the analysis of their algorithm is tight.
However they show how a direct randomized LP rounding algorithm
with approximation factor 2.54 can be obtained for this problem.
Their approach has been generalized by Sharma, Swamy, and
Williamson [Y. SHARMA, C. SWAMY, AND D. P. WILLIAMSON,
Approximation algorithms for prize collecting forest problems with
submodular penalty functions, in Proceedings of the eighteenth
annual ACM-SIAM symposium on Discrete algorithms (SODA '07),
Philadelphia, Pa., USA, 2007, Society for Industrial and Applied
Mathematics, pp. 1275-1284] for network design problems where
violated arbitrary 0-1 connectivity constraints are allowed in
exchange for a very general penalty function. The work of
Hajiaghayi and Jain has also motivated a game-theoretic version of
the problem considered by Gupta et al. [A. GUPTA, J. K{umlaut over
( )}ONEMANN, S. LEONARDI, R. RAVI, AND G. SCH{umlaut over ( )}AFER,
An efficient cost-sharing mechanism for the prize-collecting
Steiner forest problem, in Proceedings of the eighteenth annual
ACM-SIAM symposium on Discrete algorithms (SODA '07), Philadelphia,
Pa., USA, 2007, Society for Industrial and Applied Mathematics, pp.
1153-1162].
[0011] Besides the network design criteria that are able to be
considered by the PCSF problem, it may be useful to be able to
impose a further network design consideration, namely that of
providing a desired redundant connectivity between required points
(measured in terms of numbers of required alternate edge disjoint
paths between the points). For example, in order to provide for
continued service in the event of a failure of a selected network
element, such as breakage of a fiber optic cable, it may be
required that selected points (vertices) be connected with two or
more redundant edge disjoint paths (commonly referred to a
k-coverage when the design is to include k redundant paths between
points). An example of a method for finding such independent edge
disjoint paths in a network context is shown in U.S. Pat. No.
6,928,484.
[0012] Finding a network whose design considerations include a
minimization of cost, the selection of elements with assigned
costs, the requirement that the solution include pairs of selected
points, the requirement of multiple connectivity between selected
pairs of points and the allowance of assigned penalty costs if a
required pair of points is not connected by the required number of
elements, is a design problem corresponding mathematically to the
Prize Collecting Survivable Steiner Network (PCSSN) problem.
Mathematically, the network of possible elements is a graph G=(V,E)
where V is the set of vertices v.sub.i, v.sub.j, . . . and E is the
set of edges e.sub.ij, with a set of required pairs
={(s.sub.1,t.sub.1), (s.sub.2,t.sub.2), . . . (st}, a non-negative
cost function c: E.fwdarw.Q+, connectivity requirements r.sub.uv
for all pairs of vertices u and v, and a non-increasing
non-negative marginal penalty function .pi.: .fwdarw.Q+ for u and v
in case we cannot satisfy all r.sub.uv. Each edge has a cost
c.sub.ij and each pair has a penalty .pi..sub.k that is incurred if
the pair vertices s.sub.k,t.sub.k are not all connected by edges in
accordance with connectivity requirement r.sub.uv. The goal is to
find a minimum way of constructing a network (graph) H in which we
connect u and v with r'.sub.u v.ltoreq.r.sub.uv edge-disjoint paths
and pay the marginal penalty for each of the r.sub.uv-r.sub.uv
occurrences of violated connectivity between u and v.
[0013] A simple example of such a network is shown in FIG. 2. As in
FIG. 1, the graph G of potential network elements includes hollow
circle vertices that are denoted with upper case letters A, B, C,
D, F and J and represent cities, for example. The solid circles
denoted with lower case letters a, b, c, d and f represent non-city
vertices. Edges e representing transmission links, for example,
that are available to be included in the network are denoted by
solid lines linking vertices as shown. Each edge e.sub.ij has a
cost c.sub.ij. Required pairs of vertices , , , , representing
cities that are required to be connected in the network, have
connectivity requirements r.sub.1, r.sub.2, r.sub.3, r.sub.4. In
the example of FIG. 2, r.sub.1, r.sub.2, r.sub.3 are each 2 and
r.sub.4 is 4. Non-increasing non-negative marginal penalty costs
.pi..sub.11 and .pi..sub.21, for r.sub.1;.pi..sub.21 and
.pi..sub.22 for r.sub.2; .pi..sub.31 and .pi..sub.31 for r.sub.3;
and .pi..sub.41, .pi..sub.42, .pi..sub.43 and .pi..sub.44 for
r.sub.4 are available to be incurred for each instance in which we
fail to satisfy a required pair's connectivity requirement with
selected edges.
[0014] A minimum cost solution H to the graph G of FIG. 2 is shown
in FIG. 2A, where solid lines represent edges selected from those
available in FIG. 2, and dashed lines show penalty connections. The
2-fold connection for required pair AB is furnished by penalty
.pi..sub.11 and edges Ab, bB; for pair BC by penalty .pi..sub.21
and edges Bc, cD; for pair CD by edge CD and edges Cd, dD; and the
4-fold connection for pair DF is furnished by two penalties
.pi..sub.41, .pi..sub.42; edges DC, Cc, cJ, JF; and edges Dd, df,
fF.
[0015] When all connectivity requirements r.sub.uv are 1, the PCSSN
problem reduces to the PCSF problem described above. When all
penalties .pi. are .infin., the PCSSN problem reduces to the
special case of the classic survivable Steiner network design
problem. For this problem, Jain [K. JAIN, A factor 2 approximation
algorithm for the generalized Steiner network problem,
Combinatorica, 21 (2001), pp. 39-60], using the method of iterative
rounding, obtained a 2-approximation algorithm, improving on a long
line of earlier research that applied primal-dual methods to this
problem.
[0016] Jain's iterative rounding approach has been limited to the
non-prize collecting context of the classic problem he addresses,
with no provision for penalties, and does not provide feasible
solutions to the Prize Collecting Survivable Steiner Network
problem, which heretofore have not been found.
[0017] The chief problem encountered in finding an optimum network
design solution in context of PCSF and PCSSN is that for all but
trivial configurations, which means that for all configurations
actually encountered in the real world of network design, methods
providing exact solutions may be devised but they are not feasible
because they require too much time to compute. For example, it
would be possible to obtain an exact solution by making a brute
force comparison of the costs of all possible network element
permutations but years of computation time would be needed for
routinely complex networks. Accordingly, methods providing
approximate solutions have been sought. Approximation methods have,
for the most part, used linear program solving techniques to solve
optimization problems in which the variables correspond to the
physical attributes (locations, connection cost, penalties,
connectivity) of the network elements. However, approximation
methods often must make limiting assumptions in order to be
solvable using known ellipsoidal or flow-based LP-solver
techniques. Additionally, approximation algorithms must provide
some measure of how closely the approximate solution is to an
optimal solution, which means that an approximation solution must
be mathematically demonstrated to be within a factor p of the
optimum solution (such an algorithm is referred to as a
p-approximation algorithm). To prove the approximation has a
specified degree of accuracy, limiting assumptions may be made that
may reduce the generality of the methods in dealing with actual
design applications.
[0018] Thus approximation methods must find a suitable balance
among three considerations: the computation time needed to perform
them, the computational power needed to perform them, and the
degree to which the result can be shown to compare to an optimum
result. Moreover, techniques that are known to be useful in some
situations, such as primal-dual solutions, or randomized rounding
solutions, cannot generally be extended in any straightforward
manner to more complicated networks' additional variables.
[0019] Accordingly, there is a need to provide an approximation
algorithm for use in network design methods that can provide
solutions when network design criteria correspond to a PCSF or a
PCSSN. There is a further need to provide such an approximation
algorithm that further has a suitable approximation factor, that
can use existing LP-solvers, and that can find solutions using
reasonable computational resources in a reasonable time.
BRIEF SUMMARY OF THE INVENTION
[0020] Briefly, the present invention is a method for designing
minimum cost networks by selecting from a collection of possible
network elements those which provide a minimum total cost given
further network design constraints. The present invention, for the
first time, uses an iterative rounding approach for
prize-collecting versions of Steiner forest problems and more
generally to survivable Steiner network design. To the best of our
knowledge, this method of iterative rounding has not been used in a
prize-collecting context, even for the classic prize-collecting
Steiner tree.
[0021] In a first exemplary embodiment, the iterative rounding
algorithm is used in a method to find minimum cost networks by
selecting network elements with assigned costs where the design
requires either the connection of network elements to required
pairs of nodes (vertices) or alternatively the payment of penalty
costs for failure to connect any node of a pair to a network
element (edge). The first embodiment creates a PCSF problem
incorporating the network cost and other constraints, finds an
approximate solution to the PCSF problem using the iterative
rounding algorithm, and then uses elements in the solution in the
network design. In a second embodiment, the iterative rounding
algorithm is used in a method to find minimum cost networks by
selecting network elements with assigned costs where the design
requires a specified connectivity of network elements between pairs
of nodes (vertices) or the payment of penalty costs for failure to
connect any node pair with network elements to the required degree
of connectivity. The second embodiment creates a PCSSN problem
incorporating the network cost and other constraints, finds an
approximate solution to the PCSSN problem using the iterative
rounding algorithm, and then uses elements in the solution in the
network design.
[0022] In a specific embodiment, the iterative rounding algorithm
for finding an approximate solution to a minimum cost network
according to the invention puts the network element parameters of
edge costs, required pairs, and penalties in linear program form,
obtains a first feasible solution to the linear program in terms of
element coefficients, sets a threshold .alpha. for the coefficients
in the solution, pays penalties for pairs with coefficients
exceeding the threshold .alpha. and includes the paid penalties in
the network solution, selects edges with coefficients exceeding the
threshold .alpha. and includes the edges in the network solution,
constructs a residual graph including all the edges and the pairs
remaining after removing penalties, sets coefficient values of
selected edges to 1 in the residual graph, obtains a feasible
solution to a linear program for the residual graph, and repeats
these steps until it is determined that all required pairs
connectivity requirements have been met either by connection to
selected edge elements or by payment of non-connection penalties.
The solution obtained by this iterative rounding process can be
proven to be a 3-approximation solution and it is preferable to
select .alpha.=1/3.
[0023] In another embodiment of the invention, instead of selecting
all pair penalties and edges whose coefficients exceed the
threshold .alpha., the method proceeds by choosing in each
iteration the pair penalty with the maximum coefficient value and
the edge with the maximum coefficient value, and then proceeding as
described above. By removing in each iteration only the maximum
coefficient valued penalties and edges, the method will find
solutions that are nearer to optimal solutions, but at the cost of
increased processing time.
[0024] The iterative rounding algorithm according to the invention
provides a solution to PCSF and PCSSN problems heretofore
unavailable. The solution is a 3-approximation algorithm, which
means that the solution in the worst case is within a factor of 3
of optimal. In practice, the solutions found by the algorithm can
be exact. Thus a number of practical network design problems, such
as designing a minimal cost telecommunications fiber optic network,
or a minimal cost package delivery network, can be solved using
methods that use readily available LP-solvers and steps that are
easily put into computer program form, to be executed by standard
computer systems. The invention advantageously provides a solution
with greater generality than previous solutions, is simple in form
and thus easily implemented and able to be performed without
exorbitant demands in terms of time and equipment.
[0025] In other aspects, the invention includes network design
systems operating in accordance with the invention, and computer
readable program means for causing a computer to perform the
methods of the invention. In a further aspect, the invention
includes networks with elements selected according to the methods
of the invention.
[0026] These and other objects, aspects, advantages and features of
the invention are set forth in the attached description.
BRIEF DESCRIPTION OF THE DRAWINGS
[0027] The foregoing summary of the invention, as well as the
following detailed description of the preferred embodiments, is
better understood when read in conjunction with the accompanying
drawings, which are included by way of example and not by way of
limitation with regard to the claimed invention:
[0028] FIG. 1 shows a network graph of a type corresponding to a
PCSF.
[0029] FIG. 1A shows a minimum-cost solution of the network graph
of FIG. 1, of a type which may be found using an iterative rounding
method according to the invention.
[0030] FIG. 1B shows a residual graph used by an iterative rounding
method according to the invention to find the minimum-cost solution
of FIG. 1A.
[0031] FIG. 2 is a network graph of a type corresponding to a
PCSSN.
[0032] FIG. 2A shows a minimum-cost solution of the network graph
of FIG. 2, of a type which may be found using an iterative rounding
method according to the invention.
[0033] FIG. 2B shows a residual graph used by an iterative rounding
method according to the invention to find the minimum-cost solution
of FIG. 2A.
[0034] FIGS. 3A and 3B form a flowchart showing a method for
designing a network according to the invention.
[0035] FIGS. 4A and 4B are a flowchart of a second method for
designing a network according to the invention.
[0036] FIG. 5 is a flowchart of a modified method for designing a
network according to the invention.
[0037] FIG. 6 is a diagram of a computer system to operate a
computer readable program means that performs the steps in the
methods of the present invention.
DETAILED DESCRIPTION OF THE INVENTION
[0038] FIGS. 3A and 3B together form a flowchart showing a method
100 for designing a network according to the invention. Method 100
may be thought of as a computer routine, or a calculation module to
be operated by a computer. Method 100 finds a solution to a PCSF
whose variables correspond to the costs, required locations and
penalties associated with the elements of a physical network to be
designed. Thus method 100 will operate on a set of possible network
elements as shown in FIG. 1 and will find a minimum cost network as
shown in FIG. 1A.
[0039] As discussed above, the desired network includes elements
selected from a set of possible elements corresponding to graph
edges, each of which has a cost. In a fiber optic network design,
for example, the edge elements will be point-to-point routes (e.g.,
city to city) and the costs may be construction costs such as those
that may be encountered in installing fiber optic cable between
city A and city B. In step 110, the method 100 assigns these
available network elements costs c.sub.e to edges e.
[0040] Also as discussed above, the network solution is to include
certain required pairs, which correspond to graph vertices. In a
fiber optic network design, for example, the required pairs may
represent pairs of cities (New York and Washington, New York and
Philadelphia, Philadelphia and Richmond, etc,) that must be
included in any solution. In step 120, the method identifies each
of these required pairs .
[0041] In step 130, the method assigns penalty costs .pi. to the
required pairs that will be incurred if selected network elements
do not connect them. In a fiber optic network design, for example,
the penalty costs may represent the cost of leasing transmission
capacity from a competing carrier in order to connect the two
cities of the required pair.
[0042] In step 140, the method constructs a natural LP for the
optimization problem. As an example, the traditional LP relaxation
for the PCSF problem which can be solved using the Ellipsoid
algorithm is as follows:
OPT = minimize e .di-elect cons. E c e x e + i , j .di-elect cons.
P .pi. ij z ij ##EQU00001## subject to e .di-elect cons.
.differential. ( S ) x e + z ij .gtoreq. 1 .A-inverted. S V , ( i ,
j ) .di-elect cons. P , S ( i , j ) ##EQU00001.2## x e .gtoreq. 0
.A-inverted. e .di-elect cons. E z ij .gtoreq. 0 .A-inverted. ( i ,
j ) .di-elect cons. P ##EQU00001.3## Here for a set S V , we denote
{ i , j } S = 1 by S ( i , j ) . ##EQU00001.4##
[0043] In this LP x.sub.e and z.sub.ij are variable coefficients of
the c.sub.e and .pi..sub.ij to be solved for in the minimum
solution. The notation .delta.(S)=the set of all edges that are cut
by cut (S: S).
[0044] The LP relaxation written as the cut-based LP set forth in
line (1) above alternatively may be written as a standard
flow-based LP in which case one would use LP-solver algorithms for
a polynomial number of variables and constraints. The choice of
which form of LP to use will depend on practical considerations,
such as which form is easier to solve using available LP-solver
software.
[0045] Referring again to FIG. 3A, the method proceeds in step 150
by finding an optimal basic solution x*, z* to the LP of step 140.
Such a solution may be obtained by using an LP-solver such as the
CPLEX.RTM. solver provided by ILOG, Inc. of Sunnyvale, Calif., USA.
In such an optimal solution, the coefficients x*, z* will be
non-integer fractional numbers, which do not provide a workable
network solution as it is not possible to create a real network out
of fractional edges or fractional penalties.
[0046] Using the optimal basic solution x*, z* to the LP of step
140, in step 160 the method then selects a number .alpha. to be
used as a threshold, 0<.alpha..ltoreq.1, and then identifies
E.sub..alpha., as the set of edges whose value in x* is at least
the threshold .alpha. and identifies P.sub..alpha. as the set of
pairs whose value in z* is at least the threshold .alpha.. As will
be explained below, it can be shown that when .alpha.=1/3, an
iterative rounding algorithm in accordance with the steps of method
100 will produce a minimum-cost 3-approximation algorithm solution.
Accordingly, it is advantageous to select .alpha.=1/3. Use of a
lower threshold, such as 1/4, would produce a 4-approximation
algorithm; use of a higher threshold, such as 1/2, would not assure
that all iterations would advance to a solution.
[0047] Next in step 162, for the set P.sub..alpha. we pay all the
penalties .pi..sub.ij of pairs (i, j) whose
z*.sub.ij.gtoreq..alpha. and include these penalties in a
minimum-cost network final solution S.sub.PCSF. (It will be
understood that when we say we "pay" the penalties, we do not
actually pay the penalty at this design stage of a network but
signify that we choose to pay the penalty when the network is
realized. For convenience, however, we shall refer to payment of a
penalty as including designating a certain penalty to be paid
later.) We may now remove these pairs with paid penalties from
further consideration in the succeeding steps of method 100 to
select other network elements for the solution.
[0048] In step 164, for the set E.sub..alpha. we select all edges e
whose x*.sub.e.gtoreq..alpha. and include them in the minimum-cost
network final solution S.sub.PCSF.
[0049] In step 170 the method forms another graph G.sub.res as
follows: (1) we remove any pairs for which we paid a penalty in
step 162, but retain the remaining pairs; (2) we include all edges
in E, including the edges E.sub..alpha.; (3) however, we fix all
values of x.sub.e of edges in E.sub..alpha. to be 1. FIG. 1B shows
an example of a graph G.sub.res for the graph of FIG. 1, where P1
has had penalty .pi..sub.1 paid, and edges Bc and cC, shown by
heavy lines, have been selected. We show x.sub.Bc and x.sub.cC both
equal to 1 but the x.sub.e for the remaining edges remain as
variables.
[0050] In step 172 we then prepare a residual LP for graph
G.sub.res, referred to as LP.sub.res, with a new OPT.sub.res as
follows, where we define E.sub.res=E-E.sub..alpha. and =-:
OPT res = minimize e .di-elect cons. E c e x e + i , j .di-elect
cons. P res .pi. ij z ij ##EQU00002## subject to e .di-elect cons.
.differential. ( S ) x e + k = 1 r ij z ij k .gtoreq. r ij
.A-inverted. S .di-elect cons. V , ( i , j ) .di-elect cons. P res
, S ( i , j ) , .differential. ( S ) E .alpha. = 0 ##EQU00002.2## x
e .gtoreq. 0 .A-inverted. e .di-elect cons. E res z ij .gtoreq. 0
.A-inverted. ( i , j ) .di-elect cons. P res x e = 1 .A-inverted. e
.di-elect cons. E .alpha. ##EQU00002.3##
[0051] Note in LP.sub.res that by including edges in E.sub..alpha.
with x.sub.e=1 and ignoring further consideration of pairs in we
always produce another LP similar in form to that for OPT in step
140. Indeed, the only differences between OPT and OPT.sub.res lie
in their constraints, which recognize that the new optimization
problem concerns a subset of the original graph G. It will be
understood by those of skill in the art that the step of
constructing a residual graph may include a step of contracting
those selected edges whose coefficients are set to 1 in the
residual graph. Moreover, if LP were written as a flow-based LP,
the new LP.sub.res would be similar in form to that LP.
[0052] Next, in step 180, the method finds an optimal basic
feasible solution x.sup.1, z.sup.1 to LP.sub.res using an LP solver
such as the CPLEX.RTM. solver referred to above.
[0053] Using the optimal basic feasible solution x.sup.1, z.sup.1
to LP.sub.res found in step 180, in step 190 the method identifies
E.sub..alpha..sup.1 as the set of edges whose value in x.sup.1 is
at least .alpha. and identifies .sup.1 as the set of pairs whose
value in z.sup.1 is at least .alpha., in the manner described above
in step 160.
[0054] In step 192 for the pairs we pay all the penalties
.pi..sub.ij of pairs (i,j) whose z.sup.1.sub.ij.gtoreq..alpha. and
include these penalties in a minimum-cost network final solution
S.sub.PCSF.
[0055] In addition, in step 194 for the edges E.sub..alpha..sup.1
we include all edges e whose x.sup.1.sub.e.gtoreq..alpha. in the
minimum-cost network final solution S.sub.PCSF.
[0056] In step 200 we determine if any there are any pairs P of the
graph G, after removing pairs + for which penalties have been paid,
that remain unconnected by the selected edges
E.sub..alpha.+E.sub..alpha..sup.1. If yes, the method returns back
to step 170 via path 210 to form a new G.sub.res.sup.1 in which
define E.sub.res.sup.1=E-E.sub..alpha.-E.sub..alpha..sup.1 and we
define .sup.1=--, we fix all values of x.sub.e of edges in
E.sub..alpha. and E.sub..alpha..sup.1 to be 1 and then loop
iteratively through steps 180 through 200 until no original pairs
of the graph G remain unconnected by selected edges or have not
been accounted for by paying a penalty. If it is determined that no
such original pairs of the graph G remain, i.e., that all pair
connectivity requirements have been satisfied, the iteration stops
and the minimum cost network solution S.sub.PCSF to OPT is obtained
as the union of the pairs .sup.1, . . . for which the penalties
.pi..sub.ij have been paid and the edges E.sub.a, E.sub.a.sup.1, .
. . which were selected to be included. FIG. 1A shows such a
minimum cost network obtained as a result of the iterative rounding
steps of method 100 through successive versions of G.sub.res, one
of which is shown in FIG. 1B. The cost of the network is then the
sum of the penalties for the pairs .sup.1, . . . and the edge costs
c.sub.ij for the edges E.sub..alpha., E.sub..alpha..sup.1, . . .
.
[0057] Then, in step 220, the solution S.sub.PCSF is used to select
network elements, and penalty alternatives, for the physical
network that method 100 is used to design. The selected network
elements may include all the penalty and edge elements identified
by method 100, or may choose only a subset of those elements if
other design requirements are present. In the fiber optic example
given previously, these network elements will be the fiber optic
links that provide connections between cities, and the penalty
alternatives will be the lease costs of providing service through
other providers to cities unconnected by the selected fiber optic
links.
[0058] It can be proven that the iterative rounding method 100
provides a 3-approximation feasible solution to OPT, and thus to
PCSF, and enables a minimum cost network to be designed with a cost
solution that is at most a factor of three greater than the optimum
cost. While the proof establishes there is a worst case bound of
three, in practice the algorithm seems to perform much better than
this, and often will produce the optimal solution or a near-optimal
solution.
[0059] The proof that method 100 provides a 3-approximation
feasible solution to OPT, which is lengthy, appears in the
Appendix.
[0060] The steps of method 100 described above may be embodied in
routines in a computer program, capable of being operated by a
standard computer system such as the one shown in FIG. 6 and
described below.
[0061] In method 400, illustrated in FIGS. 4A and 4B, the algorithm
of method 100 is extended to find a 3-approximation solution to
PCSSN. Method 400 finds a solution to a PCSSN whose variables
correspond to the costs, required pairs of locations, connectivity
requirements and penalties of components of a physical network to
be designed.
[0062] As discussed above, the network is to include elements
selected from a number of possible elements corresponding to graph
edges, each of which has a cost. In a fiber optic network design,
for example, the elements will be point-to-point routes (e.g., city
to city) and their costs may be construction costs such as those
that may be encountered in installing fiber optic cable between
city A and city B. In step 410, the method 400 assigns these
potential network elements costs c.sub.e to edges e.
[0063] Also as discussed above, the network solution is required to
include certain pairs, which correspond to graph vertices. In a
fiber optic network design, for example, the pairs may represent
pairs of cities (New York and Washington, New York and
Philadelphia, Philadelphia and Richmond, etc,) that are required to
be included in any solution. In step 420, the method identifies
each of these required pairs of vertices v.sub.i, v.sub.j.
[0064] Additionally as discussed above the network solution has
connectivity requirements r.sub.uv for the vertices u, v of a pair
. In step 430, the method assigns connectivity requirements
r.sub.uv for each of the pairs . In a fiber optic network design,
for example, the connectivity requirements may be that a pair of
cities shall have k independent (or edge disjoint) multiple paths
connecting them to protect against loss of service in the event of
element failure.
[0065] In step 432, the method assigns non-increasing marginal
penalty costs .pi..sub.uv to the pairs that will be incurred if the
resulting network does not have the required connectivity r.sub.uv
for the pairs. The marginal penalty cost will be assessed for each
connectivity violation; thus if the connectivity requirement is for
3 independent connections, and the network provides only one, then
two penalty costs will be incurred. In a fiber optic network
design, for example, each penalty cost may represent the cost of
leasing one transmission link from a competing carrier in order to
provide one of the required connectivities between the two cities
of the pair.
[0066] In step 440, the method constructs a natural LP for the
optimization problem. In this problem, we are given connectivity
requirements r.sub.uv for all pairs of vertices u and v and a
non-increasing marginal penalty function .pi..sub.uv(.) for u and
v. Our goal is to find a minimum way of constructing a graph in
which we connect u and v with r'.sub.uv.ltoreq.r.sub.uv
edge-disjoint paths and pay the marginal penalty
.pi..sub.uv(r'.sub.uv+1)+.pi..sub.uv(r'.sub.uv+2)+ . . .
+.pi..sub.uv(r.sub.uv) for violating the connectivity between u and
v to the amount of r.sub.uv-r'.sub.uv. The LP relaxation for the
PCSSN problem which can be solved using the Ellipsoid algorithm is
as follows:
OPT = minimize e .di-elect cons. E c e x e + i , j .di-elect cons.
P k = 1 r ij .pi. ij ( k ) z ij k ##EQU00003## subject to
##EQU00003.2## e .di-elect cons. .differential. ( S ) x e + k = 1 r
ij z ij k .gtoreq. r ij .A-inverted. S V , ( i , j ) .di-elect
cons. P , S ( i , j ) 0 x e .gtoreq. 0 .A-inverted. e .di-elect
cons. E z ij k .gtoreq. 0 .A-inverted. ( i , j ) .di-elect cons. P
, 0 .ltoreq. k .ltoreq. r ij ##EQU00003.3##
[0067] The LP relaxation written as the cut-based LP set forth
above alternatively may be written as a standard flow-based LP in
which case one would use LP-solver algorithms for a polynomial
number of variables and constraints. The choice of which form of LP
to use will depend on practical considerations, such as which form
is easier to solve.
[0068] In the LP above, it is easy to see that because
.pi..sub.uv's are non-increasing, without loss of generality we can
assume
0<z.sub.uv.sup.k only if z.sub.uv.sup.k+1=1 for
1.ltoreq.k<r.sub.uv
[0069] Thus an algorithm similar to that used in method 100 can be
used.
[0070] Referring again to FIG. 4A, the method proceeds in step 450
by finding an optimal basic solution x*, z* to the LP of step 440
by using an LP-solver such as the CPLEX.RTM. solver described
above.
[0071] Using the optimal basic solution x*, z* to the LP of step
440, in step 460 the method then selects a number .alpha. as a
threshold, 0<.alpha..ltoreq.1, and then identifies E.sub..alpha.
as the set of edges whose value in x* is at least .alpha. and
identifies r.sub..alpha. as the set of pair connectivities r.sub.uv
whose value in z* is at least .alpha.. It can be shown that when
.alpha.=1/3, an iterative rounding algorithm in accordance with the
further steps of method 400 will produce a minimum-cost
3-approximation algorithm solution. Accordingly, it is advantageous
to select .alpha.=1/3.
[0072] Next in step 462, for the set r.sub..alpha. we pay all the
connectivity penalties .pi..sub.uv for those r.sub.uv whose
z*.sub.ij.gtoreq..alpha. and include these penalties in a
minimum-cost network final solution S.sub.PCSSN. (It will be
understood that when we say we "pay" the penalties, we do not
actually pay the penalty at this design stage of a network but
signify that we choose to pay the penalty when the network is
realized. For convenience, however, we shall refer to payment of a
penalty as including designating a certain penalty to be paid
later.) We may now remove these paid connectivity requirements
r.sub.uv in pairs from further consideration in the succeeding
steps of method 400 to select other network elements for the
solution.
[0073] In step 464 for the set E.sub..alpha. we select all edges e
whose X*.sub.e.gtoreq..alpha. and include them in the minimum-cost
network final solution S.sub.PCSSN.
[0074] In step 470 the method 400 forms another graph G.sub.res as
follows: (1) we remove any connectivity requirements r.sub.uv in
pairs for which we paid a penalty in step 462, but retain the
remaining connectivity requirements in any pairs which have not had
all their connectivity requirements satisfied; (2) we include all
edges in E, including the edges E.sub..alpha.; (3) however, we fix
all values of x.sub.e of edges in E.sub..alpha. to be 1. FIG. 2B
shows an example of a graph G.sub.res for the graph of FIG. 2,
where and have had penalties .pi..sub.11 and .pi..sub.21 paid and
have possible connectivity penalties .pi..sub.12 and .pi..sub.22
remaining, and have all their possible connectivity penalties
remaining, and edges CD, Dd and cC, shown by heavy lines, have been
selected. We show x.sub.CD, x.sub.Dd and x.sub.cC all equal to 1
but the x.sub.e for the remaining edges remain as variables.
[0075] In step 472 we then prepare a residual LP for graph
G.sub.res , referred to as LP.sub.res, with a new OPT.sub.res as
follows, where we define E.sub.res=E-E.sub..alpha., and as the
pairs with remaining r.sub.uv-r.sub..alpha. connectivity
requirements that have not been paid.
OPT res = minimize e .di-elect cons. E c e x e + i , j .di-elect
cons. P res k = 1 r ij .pi. ij ( k ) z ij k ##EQU00004## subject to
##EQU00004.2## e .di-elect cons. .differential. ( S ) x e + k = 1 r
ij z ij k .gtoreq. r ij .A-inverted. S V , ( i , j ) .di-elect
cons. P res , S ( i , j ) , .differential. ( S ) E .alpha. = 0 x e
.gtoreq. 0 .A-inverted. e .di-elect cons. E res z ij k .gtoreq. 0
.A-inverted. ( i , j ) .di-elect cons. P res , 0 .ltoreq. k
.ltoreq. r ij x e = 1 .A-inverted. e .di-elect cons. E .alpha.
##EQU00004.3##
[0076] Note in LP.sub.res that by including all edges in
E.sub..alpha. with x*.sub.e.gtoreq..alpha. and ignoring
connectivity requirements r.sub.uv with paid penalties in the pairs
we can always produce another LP similar in form to that for OPT in
step 440. Indeed, OPT and OPT.sub.res are identical except for
their constraints, which reflect optimization in OPT.sub.res over a
subset of original graph G. (If LP were written as a flow-based LP,
the new LP.sub.res would be similar in form to that LP.)
[0077] Next, in step 480, the method finds an optimal basic
feasible solution x.sup.1, z.sup.1 to LP.sub.res with an LP-solver
such as the CPLEX.RTM. solver described above.
[0078] Using the optimal basic feasible solution x.sup.1, z.sup.1
to LP.sub.res found in step 480, in step 490 the method identifies
E.sub..alpha..sup.1 as the set of edges whose value in x.sup.1 is
at least .alpha. and identifies r.sub..alpha..sup.1 as the set of
pair connectivities whose value in z.sup.1 is at least .alpha., in
the manner described above in step 460.
[0079] In step 492 for the pair connectivities r.sub..alpha..sup.1
we pay all the penalties .pi..sub.ij of pairs (i,j) whose
z.sup.1.sub.ij.gtoreq..alpha. and include these penalties in the
minimum-cost network final solution S.sub.PCSSN.
[0080] In addition, in step 494 for the edges E.sub..alpha..sup.1
we include all edges e whose x.sup.1.sub.e.gtoreq..alpha. in the
minimum-cost network final solution S.sub.PCSSN.
[0081] In step 500 we determine if there are any remaining
connectivity requirements r.sub.uv in pairs of the graph G, after
removing connectivity requirements
r.sub..alpha.+r.sub..alpha..sup.1 for which penalties have been
paid, that remain unsatisfied by the included edges
E.sub..alpha.+E.sub..alpha..sup.1. If yes, the method returns back
to step 470 via path 510 to form a new G.sub.res.sup.1 with a
LP.sub.res.sup.1 in which we define
E.sub.res.sup.1=E-E.sub..alpha.-E.sub..alpha..sup.1 and we define
PG.sub.res.sup.1 as the pairs with remaining
r.sub.uv-r.sub..alpha.r.sub..alpha..sup.1 connectivity requirements
that have not been paid, we fix all values of x.sub.e of edges in
E.sub..alpha. and E.sub..alpha..sup.1 to be 1 and then we loop
iteratively through steps 480 through 500 until no original
required connectivities of the graph G remain unsatisfied by
connections with selected edges or penalty payments. If it is
determined that no unsatisfied pair connectivities of the graph G
remain, the iteration stops and the minimum cost network solution
S.sub.PCSSN to OPT is obtained as the union of the penalties
.pi..sub.ij paid for pair connectivities r.sub..alpha.,
r.sub..alpha..sup.1, . . . and the selected edges E.sub..alpha.,
E.sub..alpha..sup.1, . . . which were included in solution
S.sub.PCSSN. FIG. 2A shows such a minimum cost network obtained as
a result of the iterative rounding steps of method 400 through
successive versions of G.sub.res, one of which is shown in FIG. 2B.
The cost of the network is then the sum of the penalties
.pi..sub.ij paid for pair connectivities r.sub..alpha.,
r.sub..alpha..sup.1, . . . and the edge costs c.sub.ij for the
selected edges E.sub..alpha., E.sub..alpha..sup.1, . . . .
[0082] Then, in step 520, selected edges and penalty alternatives
in the solution S.sub.PCSSN from step 500 are used to select
network elements for the minimum cost physical network that method
400 is used to design. The selected network elements may include
all the elements identified by method 400, or may choose only a
subset of those elements if other design requirements are present.
In the fiber optic example given previously, these network elements
will be the fiber optic links that provide connections between
cities, and the penalty alternatives will be the lease costs of
providing service through other providers to pairs of cities
unconnected with the required connectivity by the selected fiber
optic links.
[0083] It will be apparent to those of skill in the art that the
method 400 becomes the same as method 100 if the connectivity
requirements r.sub.uv for pairs are all 1. However, method 400 will
permit solutions to more complex problems where at least some of
the connectivity requirements are greater than 1.
[0084] The steps of method 400 described above may be embodied in
routines in a computer program, capable of being operated by a
standard computer system such as the one shown in FIG. 6 and
described below.
[0085] FIG. 5 is a flowchart showing a modified method 550 that may
be used in methods 100 and 400 as a substitute for steps 160
through 164, 190 through 194, 460 through 464, and 490 through 494.
Briefly, in method 550, instead of selecting a threshold number a
and including all penalties and edges in the solution that have
coefficients x and z.gtoreq..alpha., method 550 proceeds by
selecting in each iteration only the penalty and edge with the
maximum coefficient x and z. Thus, in step 558 the optimal basic
solution x.sup.n, z.sup.n to an LP is received from steps 150, 180,
460 or 480, and in step 560 the maximum values of the coefficients
x.sup.max and z.sup.max in the optimal solution are found. In step
562 the penalty .pi..sub.ij whose pair connectivity requirement
corresponds to z.sup.max is paid and included in the network
solution. In step 564 the edge e.sub.ij corresponding to x.sup.max
is selected and included in the network solution. In step 556
method 550 then connects to step 170, 200, 470 or 500 in methods
100 and 400 to proceed as discussed to form a residual graph with
the connectivity penalty removed and the value of the selected edge
set to 1 and to determine if any connectivity requirements remain
unsatisfied. Methods 100 and 400 incorporating method 550 by select
only one element per iteration, rather than all which exceed a
threshold a, will require longer to perform but will have a greater
likelihood of producing a near-optimal solution.
[0086] FIG. 6 is a block diagram illustrating a computer system 600
configured to utilize iterative rounding to find an approximate
solution to the PCSF and PCSSN network design methodologies, in
accordance with the exemplary embodiments described above. Such a
computer system 600 may be utilized to implement the calculation
methods described above in regard to FIGS. 3A, 3B, 4a and 4B. Thus
the present invention, in another aspect, includes one or more
computer systems capable of carrying out the methods of the
invention; in a further aspect, the present invention is directed
to a computer-usable program code storage medium to cause a
computer to perform the methods of the invention. After reading
this description, it will become apparent to a person skilled in
the relevant art how to implement the invention using other
computer systems and/or computer architectures.
[0087] The computer system 600 shown in FIG. 6 includes one or more
processors, such as processor 604. The processor 604 is connected
to a communication bus 606. Computer system 600 also includes a
main memory 608, preferably random access memory (RAM), and can
also include a secondary memory 610. The secondary memory 610 can
include, for example, a hard disk drive 612 and/or a removable
storage drive 614, representing a floppy disk drive, a magnetic
tape drive, an optical disk drive, etc. The removable storage drive
614 reads from and/or writes to a removable storage unit 618 in a
well known manner. Removable storage unit 618 represents a floppy
disk, magnetic tape, optical disk, or similar device which is read
by and written to by removable storage drive 614. As will be
appreciated, the removable storage unit 618 includes a computer
usable storage medium for storing computer software and/or
data.
[0088] In alternative embodiments, secondary memory 610 may include
other similar means for allowing computer programs or other
instructions to be loaded into computer system 600. Such means can
include, for example, a removable storage unit 622 and an interface
620. Examples of such include a program cartridge and cartridge
interface (such as that found in video game devices), a removable
memory chip (such as an EPROM, or PROM) and associated socket, and
other removable storage units 622 and interfaces 620 which allow
software and data to be transferred from the removable storage unit
618 to computer system 600.
[0089] Computer system 600 can also include a communications
interface 624. Communications interface 624 allows software and
data to be transferred between computer system 600 and external
devices, such as an ISP network. Examples of communications
interface 624 can include a modem, a network interface (such as an
Ethernet card), a communications port, a PCMCIA slot and card, etc.
Software and data transferred via communications interface 624 are
in the form of signals which can be electronic, electromagnetic,
optical or other signals capable of being received by
communications interface 624. These signals 626 are provided to
communications interface 624 via a channel 628. This channel 628
carries signals 626 and can be implemented using wire or cable,
fiber optics, a phone line, a cellular phone link, an RF link and
other communications channels.
[0090] In this specification, the terms "computer program medium"
and "computer usable medium" are used to generally refer to media
such as removable storage device 618, a hard disk installed in hard
disk drive 612, and signals 626. These computer program products
are means for providing software to computer system 600.
[0091] Computer programs (also called computer control logic) are
stored in main memory 608 and/or secondary memory 610. Computer
programs can also be received via communications interface 624.
Such computer programs, when executed, enable the computer system
600 to perform the features of methods 100 and 400 of the present
invention as discussed herein. In particular, the computer
programs, when executed, enable the processor 604 to perform the
features of the present invention. Accordingly, such computer
programs represent controllers of the computer system 600.
[0092] In an embodiment where the invention is implemented using
software, the software may be stored in a computer program product
and loaded into computer system 600 using removable storage drive
614, hard drive 612 or communications interface 624. The control
logic (software), when executed by the processor 604, causes the
processor 604 to perform the methods of the invention as described
above.
[0093] In another embodiment, the invention is implemented
primarily in hardware using, for example, hardware components such
as application specific integrated circuits (ASICs). Implementation
of the hardware state machine so as to perform the functions
described herein will be apparent to persons skilled in the
relevant art(s). It will be apparent that the invention may be
implemented using a combination of both hardware and software.
[0094] Thus, the invention describes methods with an iterative
rounding feature enabling networks to be designed with constraints
that correspond to the PCSF and PCSSN. The feature includes both
method aspects and system aspects. While the present invention has
been described with reference to preferred and exemplary
embodiments, it will be understood by those of ordinary skill in
the art that various changes may be made and equivalents may be
substituted for elements thereof without departing from the scope
of the invention. In addition, many modifications may be made to
adapt a particular situation to the teachings of the invention
without departing from the scope thereof. Therefore, it is intended
that the invention not be limited to the particular embodiments
disclosed, but that the invention include all embodiments falling
within the scope of the appended claims.
Appendix
Proof of a 3-Approximation Solution
LP 1 Refers to the LP in Paragraph 42 of the Description; LP 6
Refers to the LP in Paragraph 50 of the Description
[0095] Theorem 1 It any basic feasible solution for LP 1, for at
least one edge e .di-elect cons. E,
x e >= 1 3 , ##EQU00005##
or for at least one pair (i, j) .di-elect cons. P,
z ij >= 1 3 . ##EQU00006##
[0096] Theorem 2 If x.sup.1, z.sup.1 is an integral solution, to
the LP 6 with value at most
1 .alpha. OPT res , ##EQU00007##
then E.sub.x.sub.e.sub.=1 .orgate. E.sub.60,
P.sub.zij=1.sup..orgate.P.sub..alpha. is feasible solution for LP 1
with value at most
1 .alpha. OPT . ##EQU00008##
[0097] Proof: The feasibility is clear from the definitions. In
order to prove the value is at most twice OPT, since the
restriction of x*, z* to E.sub.res and P.sub.res is a feasible
solution to LP 6,: OPT.sub.res.gtoreq.OPT -.SIGMA..sub.e.di-elect
cons.E.sub..alpha.x.sub.e*c.sub.e-.SIGMA..sub.(i,j).di-elect
cons.P.sub..alpha.z.sub.ij*.pi..sub.ij and thus
1 .alpha. OPT .gtoreq. 1 .alpha. OPT res + 1 .alpha. e .di-elect
cons. E .alpha. x e * c e + 1 .alpha. ( i , j ) .di-elect cons.
.alpha. z ij * .pi. ij . ##EQU00009##
Since for e .di-elect cons. E.sub..alpha., x.sub.e.gtoreq..alpha.
and for (i, j) .di-elect cons. P.sub..alpha.,
z.sub.ij.gtoreq..alpha.,
1 .alpha. OPT .gtoreq. 1 .alpha. OPT res + e .di-elect cons. E
.alpha. c e + ( i , j ) .di-elect cons. .alpha. .pi. ij .
##EQU00010##
Since for e .di-elect cons.
[0098] e .di-elect cons. E c e x e I + i , j .di-elect cons. V .pi.
ij z ij I .ltoreq. 1 .alpha. OPT res , ##EQU00011##
we have
1 .alpha. OPT .gtoreq. e .di-elect cons. E c e x e I + e .di-elect
cons. E .alpha. c e + i , j .di-elect cons. V .pi. ij z ij I + ( i
, j ) .di-elect cons. .alpha. .pi. ij . ##EQU00012##
[0099] This finishes the proof.
[0100] By combining Theorems 1 and 2 we obtain the following
conclusion:
[0101] Corollary 3 There is an iterative rounding 3-approximation
algorithm for PCSF.
Proof of Theorem 1
[0102] In this section, we prove Theorem 1. Let x, z be a basic
feasible solution. If for an edge e, x.sub.e=1 or for a pair
(i.sup.1, j.sup.1), z.sub.ij=1, then the theorem follows. Also, if
for an edge e, x.sub.e=0, then we can assume that the edge was
never there before. This assumption does not increase the cost of
the optimum fractional solution x.sub.e. Thus we can assume that
0<x.sub.e<1 and 0.ltoreq.z.sub.e<1 for all e .di-elect
cons. E and (i,j) .di-elect cons.. Let (S, ii') be the row of the
constraint matrix corresponding to a set S.OR right.V and pair
(i,i') .di-elect cons. Let x(A, B) be the guru of all x.sub.e's,
where e has one end in A and the other end in B. We represent x(A,
A) by x(A), for ease of notation. We say a set A is tight with pair
(i, i') if A .circle-w/dot. (i, i') and x(A)+z.sub.ij=1.
[0103] Theorem 4 If A is tight with (i, i') and B is tight with (j,
j') then at least one of the following holds: [0104] 1. A-B is
tight with (i, i'), B-A is tight with (j, j') and (A, ii')+(B,
jj')=(A-B, ii')+(B-A, jj'). [0105] 2. A-B is tight with (j, j'),
B-A is tight with (i, i') and (A, ii')+(B, jj')=(A-B, jj')+(B-A,
ii'). [0106] 3. A .andgate. B is tight with (i, i') A .orgate. B is
tight with (j, j') and (A, ii')+(B, jj')=(A .andgate. B, ii')+(A
.orgate. B, jj'). [0107] 4. A .andgate. B is tight with (j, j'), A
.orgate. B is tight with (i, i') and (A, ii')+(B, jj')=(A .andgate.
B, jj')+(A .orgate. B, ii'). [0108] 5. A-B is tight with (i, i'),
B-A is tight with (i', i), A .andgate. B is tight with (j, j') A
.orgate. B is tight with (j, j') and 2(A, ii')+2(B, jj')=(A-B,
ii')+(B-A, ii') +(A .andgate. B, jj')+(A .orgate. B, jj').
[0109] Proof: The proof is by case analysis. For the ease of
notaion, if a set A is tight with pair (i, i'), we assume i
.di-elect cons. A (and thus i' A).
[0110] We consider two cases: i .di-elect cons. A-B and i .di-elect
cons. A .andgate. B. Without loss of generality, we assume in the
latter case j .di-elect cons. A .andgate. B also (otherwise we
consider j instead of i in our arguments). Because of tightness we
have:
x(A)=x(A-B, B-A)+x(A-B, A .orgate. B)+x(A .andgate. B, B-A)+x(A
.andgate. B, A .orgate. B)=1-z.sub.ii'
x(B)=x(B-A, A-B)+x(B-A, A.orgate. B)+x(A .andgate. B, A-B)+x(A
.andgate. B, A .orgate. B)=1-z.sub.jj'
[0111] Let's first start with the case in which i .di-elect cons. A
.andgate. B (and thus j .di-elect cons. A .andgate. B). In this
case i' .di-elect cons. A .orgate. B and j' .di-elect cons. A
.orgate. B. Because of the feasibility:
x(A .andgate. B)=x(A .andgate. B, A-B)+x(A .andgate. B, B-A)+x(A
.andgate. B, A .orgate. B).gtoreq.1-z.sub.ii'
x(A .orgate. B)=x(A-B, A .orgate. B)+x(A .andgate. B, A .orgate.
B)+x(B-A, A .orgate. B).gtoreq.1-z.sub.jj'
[0112] Since x(.,.).gtoreq.0, by summing up the two inequalities
above and using the equalities for x(A) and x(B), we conclude that
the inequalities should be tight, i.e.. x(A .andgate.
B)=1-z.sub.ii' and x(A .orgate. B)=1-z.sub.jj' and in addition
x(A-B, B-A)=0, i.e., (A, ii')+(B, jj')=(A .andgate. B, ii')+(A
.orgate. B, jj'). Thus we are in the case 3 of the statement of the
theorem.
[0113] Now assume that i .di-elect cons. A-B and j .di-elect cons.
B-A.
[0114] Then independent of the place of i', j', by the feasibility
of the solution we have:
x(A-B)=x(A-B, A .andgate. B)+x(A-B, B-A)+x(A-B, A .orgate.
B).gtoreq.1-z.sub.ii'
x(B-A)=x(B-A, A-B)+x(B-A, A .andgate. B)+x(B-A, A .orgate.
B).gtoreq.2-z.sub.jj'
[0115] Since x(.,.).gtoreq.0, by summing up the two inequalities
above and using the equalities for x(A) and x(B), we conclude that
the inequalities should be tight, i.e., x(A-B)=1-z.sub.ii', and
x(B-A)=1-z.sub.jj', and in addition x(A .andgate. B, A .orgate.
B)=0, i.e. (A, ii')+(B, jj')=(A-B, ii')+(B-A, jj'). Thus we are in
the case 1 of the statement of the theorem.
[0116] Finally we consider the case in which i .di-elect cons. A-B
and j .di-elect cons. A .andgate. B (and thus j .di-elect cons. A
.orgate. B).
[0117] Now if i' .di-elect cons. A .orgate. B, then by the
feasibility of the solution we have:
x(A .andgate. B)=x(A.andgate. B, A-B)+x(A .andgate. B, B-A)+x(A
.andgate. B, A .orgate. B).gtoreq.1-z.sub.jj'
x(A .orgate. B)=x(A-B, A .orgate. B)+x(A .andgate. B, A .orgate.
B)+x(B-A, A .orgate. B).gtoreq.1-z.sub.ii'
[0118] Since x(.,.).gtoreq.0, by summing up the two inequalities
above and using the equalities for x(A) and x(B), we conclude that
the inequalities: should be tight, i.e., x(A .andgate.
B)=1-z.sub.jj' and x(A .orgate. B)=1-z.sub.ii' and in addition
x(A-B, B-A)=0, i.e., (A, ii')+(B, jj')=(A .andgate. B, jj')+(A
.orgate. B, ii'). Thus we are in the case 4 of the statement of the
theorem.
[0119] Finally if i' .di-elect cons. B-A then, because of
feasibility we have
x(A-B)=x(A-B, A .andgate. B)+x(A-B, B-A)+x(A-B, A .orgate.
B).gtoreq.1-z.sub.ii'
x(A .andgate. B)=x(A .andgate. B, A-B)+x(A .andgate. B, B-A)+x(A
.andgate. B, A .orgate. B).gtoreq.1-z.sub.jj'
x(B-A)+x(B-A, A-B)+x(B-A, A .andgate. B)+x(B-A, A .orgate.
B).gtoreq.1-z.sub.jj'
x(A .andgate. B)=x(A-B, A .orgate. B)+x(A .andgate. B, A .orgate.
B)+x(B-A, A .orgate. B).gtoreq.1-z.sub.ii'
[0120] Since x(.,.).gtoreq.0, by summing up the four inequalities
above and and use the equalities for 2x(A) and 2x(B), we conclude
that all inequalities should be tight, and in addition x(A-B,
B-A)=0 and x(A .andgate. B, A .orgate. B)=0, i.e., and 2(A,
ii')+2(B, jj')=(A-B, ii')+(B-A, jj')+(A .andgate. B, ii')+(A
.orgate. B, jj'). So the case 5 of the statement of the theorem
holds.
[0121] Note that especially Case 5 in Theorem 4 is novel to our
extension of iterative rounding methods.
[0122] Let be the set of all tight constraints. For any set of
tight constraints we denote the vector space spanned by the vectors
(S, ii'), where S .OR right. V and (i, i') .di-elect cons. , by
Span . We say two sets A and B cross if none of the sets A-B, B-A
and A .andgate. B is empty. We say a family of tight constraints is
laminar if no two sets corresponding to two constraints in it
cross.
[0123] Theorem 5 For any maximal laminar family of tight
constraints, Span(=Span).
[0124] Proof: Since .OR right. , Span() .OR right. Span(). If
Span() .OR right. Span(), then there exists a set A tight with (i,
i'), in such that (A, ii') Span(). We choose a set A which crosses
the minimum number of sets in Since (A, ii') Span(), (A, ii') From
the maximality of , A must cross some set B tight with (j, j') in
From Theorem 4, one of the following must hold:
1. A-B is tight with (i, i'), B-A is tight with (j, j') and (A,
ii')=(A-B, ii')+(B-A, jj')-(B, jj'). [0125] 2. A-B is tight with
(j, j'), B-A is tight with (i, i') and (A, ii')=(A-B, jj')+(B-A,
ii')-(B, jj'). [0126] 3. A .andgate. B is tight with (i, i'), A
.orgate. B is tight with (j, j') and (A, ii')=(A .andgate. B,
ii')+(A .orgate. B, jj')-(B, jj'). [0127] 4. A .andgate. B is tight
with (j, j'), A .orgate. B is tight with (i, i') and (A, ii')=(A
.andgate. B, jj')+(A .orgate. B,ii')-(B, jj'). [0128] 5. A-B is
tight with (i, i'), B-A is tight with (i', i), A .andgate. B is
tight with (j, j'), A .orgate. B is tight with (j, j') and
(A,ii')=1/2(A-B,ii')+(B-A,ii')+(A.andgate.B,jj')+(A.orgate.B,jj')]-(B,jj'-
), But this means one of (A-B, ii'), (A-B, jj'), (A .andgate. B,
ii'), (A .andgate. B,jj'), (B-A, ii'), (B-A, jj' (A .orgate. B,
ii'), and (A .orgate. B, jj') does not belong to Span. Now the rest
of the proof is very similar to that of Lemma 04.2 in [12] and thus
omitted.
[0129] Since x, z is a basic feasible solution, the dimension of
Span is |E(G)|+. Since Span(=Span, it is possible to choose a basis
for Span() from the vectors in {M(S, ii)} .di-elect cons. Let .OR
right. forms a basis for Span(. Hence we have the following
theorem.
[0130] Corollary 6 There exists a laminar family, of tight
constraints satifying the following.
[0131] 1. |=|E(G)|+|.
[0132] 2. The vectors in are independent.
[0133] 3. All constraints in are tight.
[0134] Note that in our laminar family if a set S is tight with
both (i, i') and (j, j') in two different constraints, since
z.sub.ii'=z.sub.jj', we can remove variable z.sub.jj' and just use
z.sub.ii' instead. Since we removed one variable and one
constraint, still we have a basic feasible solution which is
laminar. By this reduction, we always can make sure that each set
is tight with only one pair. Thus a tight set uniquely determines
the tight pair and we use a tight constraint and a tight set
interchangeably in our discussion below.
[0135] Now we are ready to prove Theorem 1.
[0136] Theorem 7 In any basic feasible solution for LP 1, for at
least one, edge e .di-elect cons. E,
x e .gtoreq. 1 3 , ##EQU00013##
or for at least one pair (i, j) .di-elect cons. P;
z ij .gtoreq. 1 3 . ##EQU00014##
[0137] Proof: We are giving a token to each end-point of an edge
(and thus two tokens for an edge) and two tokens to all z variables
(notice that some z variables are used for more than one commodity
pairs as discussed above). Now, we will distribute the tokens such
that for every set in the laminar family gets at least two tokens
and every root at least four tokens unless the corresponding cut
has exactly three edges. (note that each cut has at least three
edges since the value of each variable is less than 1/3) in which
the root gets at least three token. This contradict the equality
|V(F)|=|E(G)|+ where F is the rooted forest of laminar sets in the
laminar family. The subtree of F rooted at R consists of R and all
its descendants. We will prove this result by the induction on
every rooted subtree of F.
[0138] Consider a subtree rooted at R. Since all x.sub.e and
z.sub.ij are at most 1/3, if R is a leaf node, it has at least
three edges crossing it and thus gets at least three tokens (and
more than 3 tokens if the degree is more than 3). This means the
induction is correct for a leaf node, as the basis of the
induction.
[0139] If R has four or more children, by the induction hypothesis
each child has at least three tokens and each of their descendants
gets at least two tokens. We re-assign one extra token from each
child to the node R. Thus R has at least four tokens and the
induction hypothesis is correct in this case.
[0140] If R has three children, if there is a private vertex u to
R, i.e., a vertex which is in R but not in any of its children,
then we are done (since all x.sub.e values are fractional, the
degree of u is at least two and thus can contributes at least two
extra tokens toward R). Also if one of the children has at least
four edges in its corresponding cut, but the induction hypothesis
it has at least two extra tokens to contribute toward those of R
and we are done. Finally, if R has exactly three children each with
exactly three edges in its corresponding cut, then by parity R has
an odd number of edges in its corresponding cut. If R has edges in
the its cut then the three extra token by its children suffices. If
R has seven or nine edges in the cut, then at least one of its
children has all three edges in the cut and the corresponding pair
is not satisfied. But this means all other edges than those of this
cut should be zero which is contradiction to fractional value
assumption. Now if R tight with z.sub.pp' has exactly five edges in
the cut, it should be the case that two children C.sub.1 tight with
z.sub.ii' and C.sub.3 tight with z.sub.kk' have two edges in the
cut and C.sub.2 tight with z.sub.jj' has one edges in the cut. Note
that in this case z.sub.pp'>min {z.sub.ii', z.sub.jj',
z.sub.kk'} then at least for one of z.sub.ii', z.sub.jj', and
z.sub.kk'0 all pairs should be inside R for the first time and thus
we have at least two extra tokens towards the requirement of R and
we are done. It also means that p-should be inside the child C with
min {z.sub.ii', z.sub.jj', z.sub.kk'} and it should be equal to its
corresponding z value (otherwise child C violates the condition for
z.sub.pp'). Assume that z.sub.pp'=z.sub.jj'. In this case, it is
easy to see that since C.sub.2 is tight With three edges and with
five edges, the sum of x variables of C.sub.1 and C3 in the cut R
is equal to the sum of x variables of C.sub.1 and C.sub.3 to
C.sub.2. But it means at least for one of C.sub.1 and C.sub.3, the
edge e to C.sub.2 has x.sub.e.gtoreq.x.sub.e'+x.sub.e'' where e'
and e'' are the edges in the cut R. But since
x e + x e ' + x e '' > 2 3 ##EQU00015##
(due to the face that all z variables are less than 1/3),
x e .gtoreq. 1 3 ##EQU00016##
which is a contradiction. If z.sub.pp=z.sub.ii'.ltoreq.z.sub.kk'
where z.sub.ii'<z.sub.jj'. In this case the edge from C.sub.1 to
C.sub.2 should has an x value equal to that those edges of C.sub.2
and C.sub.3 in the cut R. It means the total x value of two edge of
C.sub.3 in the cut is less than 1/3 which is a contradiction, since
the third edge has x value at least 1/3.
[0141] Now in case R tight with z.sub.pp' has two children if there
is a private vertex u to R we have at least four tokens to satisfy
R (two from u and one from each of its children). If both of these
children have degree at least four, then we have four extra tokens
for R (two from each child). Then at least one of two children,
namely C.sub.1 tight with z.sub.ii', has exactly three edges in its
corresponding cut. The other child C.sub.2 tight with z.sub.jj' has
at least three edges in the cut. Note that in this case
z.sub.pp'>min {z.sub.ii', z.sub.jj'} then at least for one of
z.sub.ii' and z.sub.jj', all pairs should be inside R for the first
time and thus we have at least two extra tokens towards the
requirement of R and we are done. It also means that p should be
inside the child C with min {z.sub.ii', z.sub.jj', z.sub.kk'} and
it should be equal to its corresponding z value (otherwise child C
violates the condition for z.sub.pp'). First assume that
z.sub.pp'=z.sub.ii'.ltoreq.z.sub.jj'. In this case it is not
possible that all three edges of C.sub.1 are in the cut R, since
then all edges of C.sub.2 are in the cut R and they are zero (since
the cut R is, already tight with z.sub.pp'=z.sub.ii'). If C.sub.1
has two edges in the cut since z.sub.ii'.ltoreq.z.sub.jj', it means
sum of the x values of the edges in the cut corresponding to R,
which has one edges from C.sub.1 and the rest are the edges of
C.sub.2 in the R cut, should be at least the sum of x values of the
cut corresponding to C.sub.2. But these means the x value of the
edge of C.sub.1 in the R cut is at least the sum of x values of the
two edges from C.sub.1 to C.sub.2. Since value of all three edges
in C.sub.1 is at least 2/3, it means the x value of the edge, of
C.sub.1 in the R cut is at least 1/3, a contradiction. In case
C.sub.1 does not have any edges in the cut R, then all edges should
go C.sub.2 which means zii, should be tight with a proper subset of
edges of C.sub.2 though we know that x values of all edges of
C.sub.2 is at most 1-z.sub.jj, a contradiction. we know even all
edges of minus those edges should be tight in R with the same
z.sub.jj' which means all edges between C.sub.1 and C.sub.2 should
be zero which is a contradiction. Next assume that
z.sub.pp'=z.sub.jj'.ltoreq.z.sub.ii'. In this case it is not
possible that all edges of C.sub.2 and thus C.sub.1 are in the cut
since all edges of C.sub.1 should have zero x value. In case there
is one edge of C.sub.1, or two edges of C.sub.1 are in the cut R,
then x value of one edge of C.sub.1 is equal to the x value of two
edges of C.sub.1 which means that edge should have x value at least
1/3 B which is a contradiction. In case C.sub.1 does not have any
edges in the cut R, then C.sub.2 minus those edges should be tight
in R with the same z.sub.jj' which means all edges between C.sub.1
and, C.sub.2 should be zero which is a contradiction.
[0142] It only remains the case in which R has only one child C. In
this, case if R and C are both tight with respect to z.sub.ii, then
since R and C are independent there is a vertex u .di-elect cons.
R-C. However, if R is tight with z.sub.ii' and C is tight with
z.sub.jj' since z.sub.ii'.noteq.z.sub.jj', these two cuts should be
different and thus again there is vertex u .di-elect cons. R-C.
Since all x.sub.e values are fractional the degree of u is at least
two and thus u gets at least two tokens. Without loss of generality
assume u is the node with maximum degree. If it has degree at least
three then we can assign at least these three private tokens of u
and at least one extra token of C to R to have the induction
hypothesis satisfied. In case u has degree two and C has at least
four edges in the cut, then we have at least two tokens from u and
two extra tokens from C to assign at least four tokens to R and
satisfy the hypothesis. The only remaining case is when u has
degree two and C has an odd number of edges in its cut. However in
this case because of parity, R should has an odd number of edges in
its cut (note that in this case, we may have some other vertices
than u of degree two in R-C.) It this odd number is three then two
tokens of u and one extra token of C satisfies the required number
of tokens for R. If there is a vertex other than u in R.sub.C it
has also two extra tokens and we are done. The only case is that u
has degree two, C has three edges and all these five edges are in
the cut corresponding to R. It means in this case R should be tight
with z.sub.ii' and C should be tight with z.sub.jj' where
z.sub.ii'<z.sub.jj' (otherwise the edges from u in the cut
should zero which is a contradiction to the fractional values for
x.sub.es). Here i.noteq.u otherwise, u has degree three and thus
three extra tokens and we have at least four tokens for R. It means
i .di-elect cons. C which is again a, contradiction since, the
current cut for C violates the cut condition for i in the LP.
[0143] Finally, it is worth mentioning though we guarantee that
during the course of the algorithm, we can get a variable Which is
only one-third-integral, in the first iteration always we can find
an integral z variable. Below there is a more general proposition
regarding this issue.
[0144] Proposition 8 If there is a set S of fractional variables
which contains exactly one variable from each tight constraint in
our laminar family, our solution cannot be a basic optimum
solution. In particular, there is no basic optimum solution in
which all constraints are tight with fractional z variables.
[0145] Proof: The second statement follows immediately by taking
set S in the first statement to be the set of all fractional z
variables. The first statement follows from the fact that we can
always increase (decrease) each variable .omega. in set S by
.epsilon.(1-.omega.), for a very small .epsilon.>0, and decrease
(increase) each other variable u by .epsilon.u (increase/decrease
is depending on which option does not increase the objective
function). It is easy to see that in this way we can always get
another feasible solution which makes all our current constraints
in the laminar family tight and whose value is not larger than that
of optimum.
* * * * *